The Experiment
We see a lot of advice in prompt engineering about assigning specific roles to our LLM friends. But how much does it really matter?
I thought financial advice would be a great place to test this, because let’s face it - you expect very different financial advice from the middle aged suburban dad next door than you do from the suited up financial advisor at your bank. And you file one somewhere deep inside the farthest most awkwardly located cabinet in your brain and you act on the other hoping it'll allow you to retire a few years earlier.
So the background of the person giving the advice clearly matters here!
The Transcript
ChatGPT
Gemini
The Analysis
ChatGPT suggested a diversified portfolio both with and without assigning it a role. The key difference between the 2 responses was
In the second response the LLM asked for the investment horizon, which is more context. So it actively sought out context.
In the second response, it touched upon the principle of compounding which it just did not in the first.
Gemini on the other hand was a surprise
On the first more generic prompt without role assignment, it gave a very simplistic distribution of the money. However, it went the extra mile really explaining the principle of compounding using interactive tools
On the second attempt with role assignment, the actual portfolio distribution was more diversified. But the output contained a lot of jargon, and did not do anything to explain those concepts. So while it seems to have ‘elevated’ its own position, if I went into it understanding nothing about finance I would actually prefer the first response.
Both ChatGPT and Gemini are basically reinforced learning models. This means that they learn due to a process that rewards positive behaviour and punishes negative ones (think training a dog). Who does this? - humans!!
The reason the models seem to have 'personalities’ is because they are graded for what’s acceptable by humans and these human graders are told what's acceptable by the creator guidelines (OpenAI or Google).
So when asked to play financial advisor and produce an output, one set of human graders could have rewarded compliance centric responses while another set could have rewarded academic excellence.
Overall though, it seems like role assignment is a double edged sword for us as a user. While on one hand you could end up with better advice, the advice could also become esoteric and not be useful anymore.
The Moat
AI will only always have as much context as you feed it. It might ask a few questions but those will mostly be shallow ones.
If you were to go to a real financial advisor (and they are good), they will ask you many many many questions. And they are not for nothing. The more they know you, the more specific advice they can give you.
In theory one could argue that you can give AI that context and it will incorporate it. But at least as of now our lifetime of memories are not transferable on a small chip which can then be fed into the LLM.
A real human has a rich life. They might have a wedding coming up, they might have a family with special needs, they might want to retire at 50 in a beach town, they might want to buy a home, they might need to support aging parents, they might want to have kids. Additionally, they also have a relationship with money - someone who has had a difficult childhood dealing with poverty is not going to deal with money the same way as children of affluence. All of this is just too much context to feed to a machine, especially when you don't know what it even needs.
A human almost intuitively processes the nuances of another person's life. They are able to ask deeper questions. And they are able to truly feel the depth of the answers. This is what sets humans apart. Empathy and context are your human moat. And until there is a magic wand that can extract strands of memories and feed them into a computer, we're always going to need a human in the loop.